Bitrix24 Team Replaces Excel Quote Calculator With an App Built by an AI Agent

Excel lives in sales departments longer than anyone would like. While one person uses a spreadsheet, everything is fine. When ten people do, each gets their own copy: one works from an old price list, another "slightly tweaked" a formula, a third gave a client a discount above policy. The manager finds out from the signed contract. As part of the Bitrix24 DevRel article series, technical writer Igor Roslyakov tested how quickly such a calculator can be replaced with an app inside the CRM if an AI agent does most of the work.

What was built and why

The scenario is a toy one, but a recognizable one: a company sells office furniture. A manager looks up products in the catalog, adds them to an order, selects a company and contact, and enters a discount. The app must:

• take current prices from the CRM catalog, not from constants in the code;
• calculate the total by uniform rules;
• cap the discount: up to 100 000 ₽, a maximum of 5%; from 100 000 to 300 000 ₽, 10%; from 300 000 ₽, 15%;
• create a deal and a commercial proposal directly in Bitrix24.

An important nuance: the threshold is determined by the order total before the discount. If a manager enters too much, the app does not silently trim the value. It rejects the input and explains which percentage is allowed.

They used the Bitrix24 Cowork/Code agent. It has two modes: "Cowork" for working with the portal's business data and "Code" for development. In the first, the agent created a test catalog section with six products, three companies, and contacts linked to them. In the second, it built the app itself and deployed it on the Bitrix24 Vibecode platform.

How the prompt is structured

The most useful part of the piece is not the calculator itself but the structure of the request. The author did not write the prompt by hand: he first asked the agent to draft a precise specification, then ran the report through ChatGPT, edited it several times, and handed it to the agent. This also saves tokens.

The specification bakes in decisions that people usually forget to spell out:

• The server is the source of truth. The total from the client is not accepted. Before creating a deal, the server re-reads prices from the catalog and recalculates the limit. Tampering with data in the browser achieves nothing.
• Identity comes from the gateway. The current employee becomes the responsible person in the deal. The ID is taken from the gateway header, not from the request body.
• Department-based access. An employee outside the permitted department gets a clear refusal, not a blank screen. Actions in the CRM run with that employee's permissions, not as the administrator.
• Idempotency. An order key is created once per calculation. A repeat click returns the deal that was already created; a parallel duplicate gets a 409.
• Hard limits. No more than 50 lines, quantity from 1 to 999, discount from 0 to 50, and no universal proxy into the portal.
• A "what not to do" list. Approvals, VAT, margin, and an audit log are postponed to the next version. This keeps the task from growing out of control.

The order of work is also worth noting: first tests for the money logic (boundaries 99 999,99 / 100 000 / 299 999,99 / 300 000), then the server, then the interface, a security check, and publication. The agent creates a test deal only after a human confirms.

Limitations and what to take away

First, some caveats. This piece comes from the DevRel team of Bitrix24 itself, so it shows the vendor's tool in a favorable light. The case is a model one: a catalog of six products and a single discount rule are not the same as a real price list with hundreds of items and exceptions. The available part of the text contains no final figures on time or the number of iterations, so the savings cannot be assessed. Questions also remain that the author himself raises separately: how safe it is to trust calculations to an app written by AI, and who will maintain it.

The practical takeaway for developers and product managers extends beyond Bitrix24. If you hand an internal tool to an AI agent, the quality of the result is determined not by the model but by the spec. Describe where the truth lives (the server), what must not be accepted from the client, how to guard against duplicates, and what is explicitly out of scope for the version. Check money and permissions with tests at value boundaries. And do not create production entities in the CRM without human confirmation: the author wrote that requirement into the spec as well.